This repository contains the LungAttn for lung sound deep learning classification model, published in this paper.
The architecture of our LungAttn model. The input is a 3-channel spectrogram after tunable-Q wavelet transform (TQWT) and short time Fourier transform (STFT)
In order to train the model, you need to download ICBHI 2017 database here. Each sample provided by this database contains several breath cycles. So you need to clip them according to the start and end time declared officialy. Then you need to divide them into train set and test set. Here we divide them based on official suggestion.
The class to clip samples and divide database are concluded in
LungAttn/pre-processing/tqwt_stft.py
named clip_cycle and clip_test respectively.
After that, we implement tunable-Q wavelet transform (TQWT) to decompose the original lung sound and short time Fourier transform (STFT) to convert the audio into spectrograms.
You can run
LungAttn/pre-processing/tqwt_stft.py
to store the spectrograms as pictures locally. Then
LungAttn/pre-processing/pm_pack.py
helps you to store spectrograms and corresponding labels into .p file.
To eliminate the imbalanced problem of ICBHI 2017 dataset, we implement mixup data augmentation method.
The implementation of mixup method is included in
LungAttn/model/LungAttn.py
named mixup_data.
The model was built using PyTorch, please read detail in
LungAttn/model/LungAttn.py
To run the model, you can use the command
python3 model/LungAttn.py \
--lr 0.1 \
--gpu 0 \
--nepochs 100 \
--input ../pack/official/tqwt1_4_train.p \
--test ../pack/official/tqwt1_4_test.p \
--batch_size 32 \
--mixup True \
> log/outfile/myout.file 2>&1&
Comparison with state-of-the art works:
Confusion matrix:
- Jizuo Li
- Jiajun Yuan
- Hansong Wang
- Shijian Liu
- Qianyu Guo
- Yi Ma
- Yongfu Li*
- Liebin Zhao*
- Guoxing Wang
Please cite these papers if you have used part of this work.
Li J, Yuan J, Wang H, et al. LungAttn: advanced lung sound classification using attention mechanism with dual TQWT and triple STFT spectrogram[J]. Physiological Measurement, 2021, 42(10): 105006.